> That causal relationship is encoded in language, it doesn't have to be inferred from i.i.d. observational data. Seems to be the emerging consensus to explain

> That causal relationship is encoded in language, it doesn't have to be inferred from i.i.d. observational data. Seems to be the emerging consensus to explain the magic of LLMs. This was addressing the theoretical impossibility for an LLM to infer causal relationship from independent and identically distributed ("i.i.d.") observations. To me this seems to state if IFF LLMs are indeed inferring causal structure (*), then there is no such thing as an "i.i.d. observation" using language. * is this actually proven given the occasional errors and misbehavior of LLMs?